Meituan Open-Sources LongCat-2.5: Long-Sequence SFT Single-GPU Throughput Up 86% vs. DDP
On November 10, Meituan open-sourced LongCat-2.5 under the MIT license, enabling commercial inference deployment and further training. The model targets long-horizon agents and is natively multimodal, with 1.6 trillion total parameters and about 48 billion activated per token. It supports a one-million-token context and significantly improves end-to-end delivery in complex codebase development, complete project construction, Visual Coding, and professional office work. The weights are available on HuggingFace, along with a technical blog and API. The capability improvements come from scaling the training pipeline.